Ngseo/openvla-oft-libero-4suite-mix-aug07-sigregconcat-step400000

OpenVLA-OFT × V-JEPA Disentangle — LIBERO 4-suite mix + aug + SIGReg(concat) (cumulative step 400000)

LoRA adapter + action_head + pooler checkpoint from Stage 3 disentangle training on the LIBERO 4-task-suite mix with two extras vs the prior step200000 release:

• Online domain randomization (aug_p=0.7, lighting+noise+crop)
• SIGReg with concat([z, y_task, y_dom]) replacing the explicit dom_decorr_loss (λ_dom_decorr=0, λ_sigreg=0.126)

Loss objective effectively:

loss = 24.4·L1_act + 2.94·InfoNCE(z_task, y_task) + 0.126·SIGReg([z; y_task; y_dom])

(λ_task_cos = 0 — cos is implicit in InfoNCE; weights chosen so initial contribution ratio is 10 : 8 : 1.)

This repo contains only the trainable delta on top of openvla/openvla-7b (LoRA r=32 + heads). Base 7B weights NOT included.

Files

File	Size	What	params
vla/adapter_model.safetensors	463 MB	PEFT LoRA adapter (r=32) on OpenVLA-7B	110.8 M
action_head.pt	289 MB	ResNetActionHead (MLPResNet 2 blocks, 8192→2048→1600)	151.1 M
pooler.pt	65 MB	DisentanglePooler (mean → 4096→2·4096)	33.6 M
vla/{tokenizer,processor,*.json}	~3 MB	tokenizer + processor configs	—
dataset_statistics.json	12 KB	action+proprio q01/q99 normalization stats	—
Total trainable			295.5 M

Training config

Base model	openvla/openvla-7b
LoRA rank / dropout	32 / 0.0
Pooler mode	mean
V-JEPA target	vit_large, proj_dim=4096, pooler_depth=2, num_queries=4, mlp_proj=False
Stage 1 V-JEPA ckpt	vitl.maniskill.256px.16f.4096d.frozen-d2q4.composed.infonce.45-45-10
Dataset	libero_4_task_suites_no_noops (spatial+object+goal+libero-10)
Sampling weights	spatial 0.194 / object 0.245 / goal 0.190 / libero-10 0.371
Future frames	8 (stride 2)
Batch size	16 (single GPU)
LR / decay	5e-5 / step decay at 100k → 5e-6
Cache	V-JEPA targets precomputed (vjepa_cache_oft_vitl_4096d_libero_v2)
Online augmentation	aug_p=0.7, lighting + sensor noise + per-edge crop
SIGReg input	concat([z, y_task, y_dom]) (sketch_dim=1024)
Loss weights	λ_act=24.4, λ_task_nce=2.94, λ_sigreg=0.126, λ_task_cos=0, λ_dom_decorr=0
Trained steps	400 000 (150K initial + 250K resumed; mid-run hang at ~150K)
Train env	single A100 80GB, num_workers=6

Provenance / resume

step 0     ── scratch  (b16-1gpu-aug07-sigregconcat-400k-nw6-save50k)
step 150K  ── ckpt → mid-run hang (data path issue)
step 150K → 400K  ── resumed from step150000 with our local dataset mirror

Loss progression (smoothed, last batches of each phase):

metric	step 50k	step 150k	step 400k
act L1	0.13	0.11	0.090
task (InfoNCE)	3.70	3.61	3.63
sig (concat)	2.74	2.59	2.31
total	7.09	6.58	6.12

Usage

# Base + adapter
huggingface-cli download openvla/openvla-7b
huggingface-cli download Ngseo/openvla-oft-libero-4suite-mix-aug07-sigregconcat-step400000

# Resume training (or evaluate)
torchrun --standalone --nproc-per-node 1 \
    openvla-oft/vla-scripts/train_stage3_oft_disentangle.py \
    --vla_path openvla/openvla-7b \
    --vjepa_arch vit_large --vjepa_proj_dim 4096 \
    --vjepa_pooler_depth 2 --vjepa_pooler_num_queries 4 \
    --pooler_mode mean --lora_rank 32 \
    --batch_size 16 --learning_rate 5e-5 \
    --dataset_name libero_4_task_suites_no_noops \
    --aug_p 0.7 \
    --lambda_action 24.4 --lambda_task_cos 0 --lambda_task_nce 2.94 \
    --lambda_dom_decorr 0 --lambda_sigreg 0.126 --sigreg_concat_target True \
    --resume_from_step_dir /path/to/this/ckpt ...

Sister checkpoints:

• Ngseo/openvla-oft-libero-spatial-disentangle-step60000 — spatial-only, no aug, dom_decorr
• Ngseo/openvla-oft-libero-4suite-mix-disentangle-step200000 — mix, no aug, dom_decorr
• this — mix + aug + SIGReg(concat) replacing dom_decorr, 400K